A Comparative Analysis of Deep Learning-Based Methods for Multivariate Time Series Imputation with Varying Missing Rates

Multivariate Time Series (MTS), pervasive in fields such as industry and healthcare, enable the understanding of time-dependent phenomena, leading to better predictions, more effective interventions, and improved decision-making across various applications. A common issue in MTS is missing data, which arises from diverse sampling mechanisms, ranging from malfunctioning sensors to medical exams conducted at different time intervals. Given that missing data in MTS can jeopardize downstream tasks such as forecasting and classification, numerous methods have been developed to address data imputation in MTS. However, a comparative analysis of these data imputation methods across varying missing rates is still needed. In this work, we compared five commonly used deep learning-based imputation methods (M-RNN, US-GAN, GP-VAE, SAITS, and BRITS) on the well-known Physionet Challenge 2012 dataset considering different percentages of missing data. Our results showed that SAITS exhibited the lowest average error while BRITS demonstrated lower error dispersion.